Making AI-Generated Feedback Matter: From Provision to Student Enactment

Synthesis: Alsaiari et al. (2026) report a large-scale quasi-experimental sequential cohort study (13,037 students; 51,296 student-authored resources; 70 course offerings) comparing three AI-mediated feedback workflows implemented in the RiPPLE platform. Students in the Enacted Feedback condition — prompted to select feedback suggestions, evaluate their relevance, and engage in targeted AI dialogue anchored to those selections — showed significantly higher uptake of AI-generated feedback (26.2% estimated probability) than Directed Feedback (14.1%) or Self-Directed Feedback (0.1%), along with higher Self-Assessment confidence and submitted-work quality. The finding positions student enactment, not comment quality, as the decisive variable in AI feedback, connecting to Feedback Loop, Feedback Literacy, Self-Regulated Learning, and Human AI Collaboration research.

From Provision to Enactment

Feedback value depends on two distinct challenges: providing high-quality, timely, individualized feedback comments at scale, and supporting students to interpret, evaluate, and act on those comments productively. Generative AI credibly addresses the provision challenge — with evidence that AI-generated comments can approach educator feedback in pedagogical quality — but students' uptake of those comments remains consistently limited without structured support. The paper frames this as a distinction between the provision of feedback comments and feedback use, the central tension of contemporary Formative Assessment and Feedback Literacy scholarship.

The Three Workflows

The study compares three theoretically distinct AI-mediated feedback workflows implemented at the platform level within RiPPLE, holding the authoring task, moderation rubric, and quality criteria constant across conditions:

• Directed Feedback (n=3,723): students received a one-time set of static AI-generated feedback comments without structured support for their use and with no dialogic pathway.
• Self-Directed Feedback (n=3,951): students could initiate optional, student-initiated AI-supported dialogue, but no feedback comments were generated and no prompt structured engagement.
• Enacted Feedback (n=5,363): students were prompted to select feedback suggestions, evaluate their relevance and applicability, and engage in targeted AI-supported dialogue anchored to those selections.

The conditions isolate specific theoretical mechanisms rather than incremental improvements to a common design: static comments (Directed), the mere availability of dialogue (Self-Directed), and structured support for feedback enactment (Enacted).

Key Findings

1. Enacted Feedback was associated with significantly higher uptake of AI-generated feedback comments, with an estimated probability of 26.2%, compared with 14.1% for Directed Feedback and 0.1% for Self-Directed Feedback.
2. Enacted Feedback was associated with significantly higher self-assessment confidence than both comparison conditions, consistent with stronger Self-Efficacy and evaluative engagement.
3. Enacted Feedback was associated with higher submitted-work quality (peer moderation scores) than both Directed and Self-Directed Feedback.
4. Merely providing access to AI-supported dialogue (Self-Directed Feedback) did not improve uptake, indicating that the availability of AI assistance, absent structural scaffolding, is insufficient to shift engagement.

Method and Platform

RiPPLE (Recommendation in Personalized Peer-Learning Environments) structures student activity across creation, review, and practice stages. Students author learning resources, moderate peers' work against a quality rubric, and practice with approved resources. Fine-grained interaction logs record every state transition between drafting, reviewing feedback comments, AI-supported dialogue, revision, and submission — enabling measures of workflow-specific uptake, revision counts, event-flow transitions, self-assessment confidence, and submitted-work quality. Analyses were conducted in R with mixed-effects models at the resource level and student ID as a random intercept. The language model varied across implementation periods (GPT-4o mini for the 2025 conditions, GPT-5 mini for 2026), while prompt structure and output format were held constant.

The design's key strength is that the Enacted Feedback workflow scaffolds three behaviors associated with Feedback Literacy: student agency through selection, evaluative judgment through prioritization, and dialogic engagement through targeted AI assistance anchored to students' chosen suggestions.

What this means for practice

• Instructors. Require students to select and justify the feedback suggestions they intend to act on before they revise, rather than distributing a comment set and leaving use to chance. The scaffolded Enacted Feedback workflow was associated with 26.2% estimated uptake of AI comments, against 14.1% for static comments and 0.1% when only optional AI dialogue was available.
• Instructors. Prompt students to judge relevance before acting on a suggestion: AI-generated comments read fluently and can appear authoritative even when they are pedagogically limited, so fluency must not be taken as a signal of quality.
• Designers. Move the design goal from generating better comments to structuring enactment — build selection, evaluative judgment, and selection-anchored dialogue into the workflow, because Scaffolding embedded in the workflow, not raw GenAI capability or optional chatbot access, drove the measured gains in confidence and work quality.
• Designers. Instrument the workflow so logs capture whether a revision followed a particular suggestion; the platform logs here established that editing occurred but did not show that an edit incorporated the suggested content.
• Researchers. Follow cohorts across several authoring cycles and into transfer tasks where AI scaffolding is withdrawn, since work quality was measured through peer moderation at a single point in each cohort's cycle and durable Feedback Literacy was not tested.

Limitations

• Sequential cohorts rather than random assignment: the three workflows ran in different semesters, with GPT-4o mini for Directed and Self-Directed Feedback (2025) and GPT-5 mini for Enacted Feedback (2026), so the model version is acknowledged as a confound alongside cohort and seasonal differences.
• Uptake was operationalized differently in each condition — an immediate move from static feedback to editing, requesting AI assistance and then editing, or any downstream edit after entering the scaffolded pathway — so the three rates are workflow-specific indicators and not a test of one identical behavior.
• Platform log data only: logs record observable actions at scale but not students' motivations or interpretations, leaving unexplained why some Enacted Feedback students bypassed suggestion selection and why Self-Directed students rarely used the optional assistance.
• No long-term learning measure: submitted-work quality came from peer moderation at a single point in each cohort's cycle, so whether the engagement benefits become durable gains in feedback literacy, evaluative judgment, or independent revision remains an open question.

Citation

Alsaiari, O., Baghaei, N., Lodge, J. M., Gašević, D., Winstone, N., & Khosravi, H. (2026). Making AI-generated feedback matter: From provision to student enactment.